Discriminative Regularization for Generative Models

TL;DR

This paper introduces a method to improve generative models by incorporating discriminative classifier representations into their training, resulting in clearer and higher-quality generated samples.

Contribution

It proposes a novel discriminative regularization technique for variational autoencoders that leverages classifier features to enhance sample quality.

Findings

Generated samples are clearer and more visually appealing.

The regularization improves the fidelity of the generated data.

Enhanced models outperform standard variational autoencoders.

Abstract

We explore the question of whether the representations learned by classifiers can be used to enhance the quality of generative models. Our conjecture is that labels correspond to characteristics of natural data which are most salient to humans: identity in faces, objects in images, and utterances in speech. We propose to take advantage of this by using the representations from discriminative classifiers to augment the objective function corresponding to a generative model. In particular we enhance the objective function of the variational autoencoder, a popular generative model, with a discriminative regularization term. We show that enhancing the objective function in this way leads to samples that are clearer and have higher visual quality than the samples from the standard variational autoencoders.